source: Papers/fopara2013/fopara13.tex @ 3418

Last change on this file since 3418 was 3417, checked in by campbell, 6 years ago

Many of the minor reviewer comments about FOPARA paper.

File size: 62.3 KB
Line 
1\documentclass{llncs}
2
3\usepackage{amsfonts}
4\usepackage{amsmath}
5\usepackage{amssymb} 
6\usepackage[english]{babel}
7\usepackage{color}
8\usepackage{fancybox}
9\usepackage{fancyvrb}
10\usepackage{graphicx}
11\usepackage[colorlinks,
12            bookmarks,bookmarksopen,bookmarksdepth=2]{hyperref}
13\usepackage{hyphenat}
14\usepackage[utf8x]{inputenc}
15\usepackage{listings}
16\usepackage{mdwlist}
17\usepackage{microtype}
18\usepackage{stmaryrd}
19\usepackage{url}
20
21\usepackage{tikz}
22\usetikzlibrary{positioning,calc,patterns,chains,shapes.geometric,scopes}
23
24%\renewcommand{\verb}{\lstinline}
25%\def\lstlanguagefiles{lst-grafite.tex}
26%\lstset{language=Grafite}
27\lstset{language=C}
28
29\newlength{\mylength}
30\newenvironment{frametxt}%
31        {\setlength{\fboxsep}{5pt}
32                \setlength{\mylength}{\linewidth}%
33                \addtolength{\mylength}{-2\fboxsep}%
34                \addtolength{\mylength}{-2\fboxrule}%
35                \Sbox
36                \minipage{\mylength}%
37                        \setlength{\abovedisplayskip}{0pt}%
38                        \setlength{\belowdisplayskip}{0pt}%
39                }%
40                {\endminipage\endSbox
41                        \[\fbox{\TheSbox}\]}
42
43\title{Certified Complexity (CerCo)\thanks{The project CerCo acknowledges the
44financial support of the Future and Emerging Technologies (FET) programme within
45the Seventh Framework Programme for Research of the European Commission, under
46FET-Open grant number: 243881}}
47\author{
48%Roberto
49R.~M. Amadio$^{4}$ \and
50%Nicolas
51N.~Ayache$^{3,4}$ \and
52%François
53F.~Bobot$^{3,4}$ \and
54%Jacob
55J.~P. Boender$^1$ \and
56%Brian
57B.~Campbell$^2$ \and
58%Ilias
59I.~Garnier$^2$ \and
60%Antoine
61A.~Madet$^4$ \and
62%James
63J.~McKinna$^2$ \and
64%Dominic
65D.~P. Mulligan$^1$ \and
66%Mauro
67M.~Piccolo$^1$ \and
68%Randy
69R.~Pollack$^2$ \and
70%Yann
71Y.~R\'egis-Gianas$^{3,4}$ \and
72%Claudio
73C.~Sacerdoti Coen$^1$ \and
74%Ian
75I.~Stark$^2$ \and
76%Paolo
77P.~Tranquilli$^1$}
78\institute{Dipartimento di Informatica - Scienza e Ingegneria, Universit\'a di
79Bologna \and
80LFCS, School of Informatics, University of Edinburgh
81\and INRIA (Team $\pi$r2 )
82\and
83Universit\`e Paris Diderot
84}
85
86\bibliographystyle{splncs03}
87
88\begin{document}
89
90\maketitle
91
92\begin{abstract}
93This paper provides an overview of the FET-Open Project CerCo
94(`Certified Complexity'). The project's main achievement is
95the development of a technique for performing static analyses of non-functional
96properties of programs (time, space) at the source level, without loss of accuracy
97and with a small, trusted code base. The main software component
98developed is a certified compiler producing cost annotations. The compiler
99translates source code to object code and produces an instrumented copy of the
100source code. This instrumentation exposes at
101the source level---and tracks precisely---the actual (non-asymptotic)
102computational cost of the input program. Untrusted invariant generators and trusted theorem provers
103are then used to compute and certify the parametric execution time of the
104code.
105\end{abstract}
106
107% ---------------------------------------------------------------------------- %
108% SECTION                                                                      %
109% ---------------------------------------------------------------------------- %
110\section{Introduction}
111
112\paragraph{Problem statement.} Computer programs can be specified with both
113functional constraints (what a program must do) and non-functional constraints
114(e.g. what resources---time, space, etc.---the program may use).  In the current
115state of the art, functional properties are verified for high-level source code
116by combining user annotations (e.g. preconditions and invariants) with a
117multitude of automated analyses (invariant generators, type systems, abstract
118interpretation, theorem proving, and so on). By contrast, non-functional properties
119are generally checked on low-level object code, but also demand information
120about high-level functional behaviour that must somehow be recreated.
121
122This situation presents several problems: 1) it can be hard to infer this
123high-level structure in the presence of compiler optimisations; 2) techniques
124working on object code are not useful in early development, yet problems
125detected later are more expensive to tackle; 3) parametric cost analysis is very
126hard: how can we reflect a cost that depends on the execution state (e.g. the
127value of a register or a carry bit) to a cost that the user can understand
128looking at source code? 4) functional analysis performed only on object code
129makes any contribution from the programmer hard, leading to less precision in
130the estimates.
131
132\paragraph{Vision and approach.} We want to reconcile functional and
133non-functional analyses: to share information and perform both at the same time
134on source code.
135%
136What has previously prevented this approach is the lack of a uniform and precise
137cost model for high-level code: 1) each statement occurrence is compiled
138differently and optimisations may change control flow; 2) the cost of an object
139code instruction may depend on the runtime state of hardware components like
140pipelines and caches, all of which are not visible in the source code.
141
142To solve the issue, we envision a new generation of compilers able to keep track
143of program structure through compilation and optimisation, and to exploit that
144information to define a cost model for source code that is precise, non-uniform,
145and accounts for runtime state. With such a source-level cost model we can
146reduce non-functional verification to the functional case and exploit the state
147of the art in automated high-level verification~\cite{survey}. The techniques
148currently used by the Worst Case Execution Time (WCET) community, which perform the analysis on object code,
149are still available but can now be coupled with an additional source-level
150analysis. Where the approach produces precise cost models too complex to reason
151about, safe approximations can be used to trade complexity with precision.
152Finally, source code analysis can be made during early development stages, when
153components have been specified but not implemented: source code modularity means
154that it is enough to specify the non-functional behaviour of unimplemented
155components.
156
157\paragraph{Contributions.} We have developed what we refer to as \emph{the labeling approach} \cite{labeling}, a
158technique to implement compilers that induce cost models on source programs by
159very lightweight tracking of code changes through compilation. We have studied
160how to formally prove the correctness of compilers implementing this technique.
161We have implemented such a compiler from C to object binaries for the 8051
162micro-controller, and verified it in an interactive theorem prover. We have
163implemented a Frama-C plugin \cite{framac} that invokes the compiler on a source
164program and uses this to generate invariants on the high-level source that
165correctly model low-level costs. Finally, the plugin certifies that the program
166respects these costs by calling automated theorem provers, a new and innovative
167technique in the field of cost analysis. As a case study, we show how the
168plugin can automatically compute and certify the exact reaction time of Lustre
169\cite{lustre} data flow programs compiled into C.
170
171\section{Project context and objectives}
172Formal methods for verification of functional properties of programs have
173now reached a level of maturity and automation that is facilitating a slow but
174increasing adoption in production environments. For safety critical code, it is
175becoming commonplace to combine rigorous software engineering methodologies and testing
176with static analysis, taking the strong points of each approach and mitigating
177their weaknesses. Of particular interest are open frameworks
178for the combination of different formal methods, where the programs can be
179progressively specified and are continuously enriched with new safety
180guarantees: every method contributes knowledge (e.g. new invariants) that
181becomes an assumption for later analysis.
182
183The scenario for the verification of non-functional properties (time spent,
184memory used, energy consumed) is bleaker and the future seems to be getting even
185worse. Most industries verify that real time systems meet their deadlines
186by simply performing many runs of the system and timing their execution,  computing the
187maximum time and adding an empirical safety margin, claiming the result to be a
188bound for the WCET of the program. Formal methods and software to statically
189analyse the WCET of programs exist, but they often produce bounds that are too
190pessimistic to be useful. Recent advancements in hardware architectures are all
191focused on the improvement of the average case performance, not the
192predictability of the worst case. Execution time is becoming increasingly
193dependent on execution history and the internal state of
194hardware components like pipelines and caches. Multi-core processors and non-uniform
195memory models are drastically reducing the possibility of performing
196static analysis in isolation, because programs are less and less time
197composable. Clock-precise hardware models are necessary for static analysis, and
198obtaining them is becoming harder as a consequence of the increased sophistication
199of hardware design.
200
201Despite the aforementioned problems, the need for reliable real time systems and programs is
202increasing, and there is increasing pressure from the research community towards
203the differentiation of hardware. The aim is to introduce alternative
204hardware with more predictable behaviour and hence more suitability for static
205analyses, for example, the decoupling of execution time from execution history
206by introducing randomisation \cite{proartis}.
207
208In the CerCo project \cite{cerco} we do not try to address this problem, optimistically
209assuming that static analysis of non-functional properties of programs will
210become feasible in the longer term. The main objective of our work is
211instead to bring together static analysis of functional and non-functional
212properties, which, according to the current state of the art, are completely
213independent activities with limited exchange of information: while the
214functional properties are verified on the source code, the analysis of
215non-functional properties is entirely performed on the object code to exploit
216clock-precise hardware models.
217
218Analysis currently takes place on object code for two main reasons.
219First, there is a lack of a uniform, precise cost model for source
220code instructions (or even basic blocks). During compilation, high level
221instructions are torn apart and reassembled in context-specific ways so that
222identifying a fragment of object code and a single high level instruction is
223infeasible. Even the control flow of the object and source code can be very
224different as a result of optimisations, for example aggressive loop
225optimisations may completely transform source level loops. Despite the lack of a uniform, compilation- and
226program-independent cost model on the source language, the literature on the
227analysis of non-asymptotic execution time on high level languages that assumes
228such a model is growing and gaining momentum. However, unless we can provide a
229replacement for such cost models, this literature's future practical impact looks
230to be minimal. Some hope has been provided by the EmBounded project \cite{embounded}, which
231compositionally compiles high-level code to a byte code that is executed by an
232emulator with guarantees on the maximal execution time spent for each byte code
233instruction. The approach provides a uniform model at the price of the model's
234precision (each cost is a pessimistic upper bound) and performance of the
235executed code (because the byte code is emulated  compositionally instead of
236performing a fully non-compositional compilation).
237%
238The second reason to perform the analysis on the object code is that bounding
239the worst case execution time of small code fragments in isolation (e.g. loop
240bodies) and then adding up the bounds yields very poor estimates because no
241knowledge on the hardware state can be assumed when executing the fragment. By
242analysing longer runs the bound obtained becomes more precise because the lack
243of knowledge on the initial state has less of an effect on longer computations.
244
245In CerCo we propose a radically new approach to the problem: we reject the idea
246of a uniform cost model and we propose that the compiler, which knows how the
247code is translated, must return the cost model for basic blocks of high level
248instructions. It must do so by keeping track of the control flow modifications
249to reverse them and by interfacing with static analysers.
250
251By embracing compilation, instead of avoiding it like EmBounded did, a CerCo
252compiler can at the same time produce efficient code and return costs that are
253as precise as the static analysis can be. Moreover, we allow our costs to be
254parametric: the cost of a block can depend on actual program data or on a
255summary of the execution history or on an approximated representation of the
256hardware state. For example, loop optimizations may assign to a loop body a cost
257that is a function of the number of iterations performed. As another example,
258the cost of a block may be a function of the vector of stalled pipeline states,
259which can be exposed in the source code and updated at each basic block exit. It
260is parametricity that allows one to analyse small code fragments without losing
261precision: in the analysis of the code fragment we do not have to ignore the
262initial hardware state. On the contrary, we can assume that we know exactly which
263state (or mode, as the WCET literature calls it) we are in.
264
265The cost of an execution is the sum of the cost of basic blocks multiplied by
266the number of times they are executed, which is a functional property of the
267program. Therefore, in order to perform (parametric) time analysis of programs,
268it is necessary to combine a cost model with control and data flow analysis.
269Current state of the art WCET technology
270\cite{stateart} performs the analysis on the object code, where the logic of the
271program is harder to reconstruct and most information available at the source
272code level has been lost. Imprecision in the analysis leads to useless bounds. To
273augment precision, the tools ask the user to provide constraints on the object
274code control flow, usually in the form of bounds on the number of iterations of
275loops or linear inequalities on them. This requires the user to manually link
276the source and object code, translating his assumptions on the source code
277(which may be wrong) to object code constraints. The task is error prone and
278hard, especially in the presence of complex compiler optimisations.
279
280The CerCo approach has the potential to dramatically improve the state of the
281art. By performing control and data flow analyses on the source code, the error
282prone translation of invariants is completely avoided. It is in fact performed
283by the compiler itself when it induces the cost model on the source language.
284Moreover, any available technique for the verification of functional properties
285can be immediately reused and multiple techniques can collaborate together to
286infer and certify cost invariants for the program. Parametric cost analysis
287becomes the default one, with non parametric bounds used as last resorts when
288trading the complexity of the analysis with its precision. \emph{A priori}, no
289technique previously used in traditional WCET is lost: they can just be applied
290at the source code level.
291
292Traditional techniques for WCET that work on object code are also affected by
293another problem: they cannot be applied before the generation of the object
294code. Functional properties can be analysed in early development stages, while
295analysis of non-functional properties may come too late to avoid expensive
296changes to the program architecture. Our approach already works in early
297development stages by axiomatically attaching costs to unimplemented components.
298
299All software used to verify properties of programs must be as bug free as
300possible. The trusted code base for verification is made by the code that needs
301to be trusted to believe that the property holds. The trusted code base of
302state-of-the-art WCET tools is very large: one needs to trust the control flow
303analyser and the linear programming libraries it uses and also the formal models
304of the hardware. In CerCo we are moving the control flow analysis to the source
305code and we are introducing a non-standard compiler too. To reduce the trusted
306code base, we implemented a prototype and a static analyser in an interactive
307theorem prover, which was used to certify that the cost computed on the source
308code is indeed the one actually spent by the hardware. Formal models of the
309hardware and of the high level source languages were also implemented in the
310interactive theorem prover. Control flow analysis on the source code has been
311obtained using invariant generators, tools to produce proof obligations from
312generated invariants and automatic theorem provers to verify the obligations. If
313the automatic provers are able to generate proof traces that can be
314independently checked, the only remaining component that enters the trusted code
315base is an off-the-shelf invariant generator which, in turn, can be proved
316correct using an interactive theorem prover. Therefore we achieve the double
317objective of allowing the use of more off-the-shelf components (e.g. provers and
318invariant generators) whilst reducing the trusted code base at the same time.
319
320%\paragraph{Summary of the CerCo objectives.} To summarize, the goal of CerCo is
321% to reconcile functional with non-functional analysis by performing them together
322% on the source code, sharing common knowledge about execution invariants. We want
323% to achieve the goal implementing a new generation of compilers that induce a
324% parametric, precise cost model for basic blocks on the source code. The compiler
325% should be certified using an interactive theorem prover to minimize the trusted
326% code base of the analysis. Once the cost model is induced, off-the-shelf tools
327% and techniques can be combined together to infer and prove parametric cost
328% bounds.
329%The long term benefits of the CerCo vision are expected to be:
330%1. the possibility to perform static analysis during early development stages
331%2.  parametric bounds made easier
332%3.  the application of off-the-shelf techniques currently unused for the
333% analysis of non-functional properties, like automated proving and type systems
334%4. simpler and safer interaction with the user, that is still asked for
335% knowledge, but on the source code, with the additional possibility of actually
336% verifying the provided knowledge
337%5. a reduced trusted code base
338%6. the increased accuracy of the bounds themselves.
339%
340%The long term success of the project is hindered by the increased complexity of
341% the static prediction of the non-functional behaviour of modern hardware. In the
342% time frame of the European contribution we have focused on the general
343% methodology and on the difficulties related to the development and certification
344% of a cost model inducing compiler.
345
346\section{The typical CerCo workflow}
347\begin{figure}[!t]
348\begin{tabular}{l@{\hspace{0.2cm}}|@{\hspace{0.2cm}}l}
349\begin{lstlisting}
350char a[] = {3, 2, 7, -4};
351char treshold = 4;
352
353int main() {
354  char j;
355  char *p = a;
356  int found = 0;
357  for (j=0; j < 4; j++) {
358    if (*p <= treshold)
359      { found++; }
360    p++;
361  }
362  return found;
363}
364\end{lstlisting}
365&
366%  $\vcenter{\includegraphics[width=7.5cm]{interaction_diagram.pdf}}$
367\begin{tikzpicture}[
368    baseline={([yshift=-.5ex]current bounding box)},
369    element/.style={draw, text width=1.6cm, on chain, text badly centered},
370    >=stealth
371    ]
372{[start chain=going below, node distance=.5cm]
373\node [element] (cerco) {CerCo\\compiler};
374\node [element] (cost)  {CerCo\\cost plugin};
375{[densely dashed]
376\node [element] (ded)   {Deductive\\platform};
377\node [element] (check) {Proof\\checker};
378}
379}
380\coordinate [left=3.5cm of cerco] (left);
381{[every node/.style={above, text width=3.5cm, text badly centered,
382                     font=\scriptsize}]
383\draw [<-] ([yshift=-1ex]cerco.north west) coordinate (t) --
384    node {C source}
385    (t-|left);
386\draw [->] (cerco) -- (cost);
387\draw [->] ([yshift=1ex]cerco.south west) coordinate (t) --
388    node {C source+cost annotations}
389    (t-|left) coordinate (cerco out);
390\draw [->] ([yshift=1ex]cost.south west) coordinate (t) --
391    node {C source+cost annotations\\+synthesized assertions}
392    (t-|left) coordinate (out);
393{[densely dashed]
394\draw [<-] ([yshift=-1ex]ded.north west) coordinate (t) --
395    node {C source+cost annotations\\+complexity assertions}
396    (t-|left) coordinate (ded in) .. controls +(-.5, 0) and +(-.5, 0) .. (out);
397\draw [->] ([yshift=1ex]ded.south west) coordinate (t) --
398    node {complexity obligations}
399    (t-|left) coordinate (out);
400\draw [<-] ([yshift=-1ex]check.north west) coordinate (t) --
401    node {complexity proof}
402    (t-|left) .. controls +(-.5, 0) and +(-.5, 0) .. (out);
403\draw [dash phase=2.5pt] (cerco out) .. controls +(-1, 0) and +(-1, 0) ..
404    (ded in);
405}}
406%% user:
407% head
408\draw (current bounding box.west) ++(-.2,.5)
409    circle (.2) ++(0,-.2) -- ++(0,-.1) coordinate (t);
410% arms
411\draw (t) -- +(-.3,-.2);
412\draw (t) -- +(.3,-.2);
413% body
414\draw (t) -- ++(0,-.4) coordinate (t);
415% legs
416\draw (t) -- +(-.2,-.6);
417\draw (t) -- +(.2,-.6);
418\end{tikzpicture}
419\end{tabular}
420\caption{On the left: code to count the array's elements
421 that are less than or equal to the treshold. On the right: CerCo's interaction
422 diagram. CerCo's components are drawn solid.}
423\label{test1}
424\end{figure}
425We illustrate the workflow we envisage (on the right of~\autoref{test1})
426on an example program (on the left of~\autoref{test1}).
427The user writes the program and feeds it to the CerCo compiler, which outputs
428an instrumented version of the same program that updates global variables
429that record the elapsed execution time and the stack space usage.
430The red lines in \autoref{itest1} are the instrumentation introduced by the
431compiler. The annotated program can then be enriched with complexity assertions
432in the style of Hoare logic, that are passed to a deductive platform (in our
433case Frama-C). We provide as a Frama-C cost plugin a simple automatic
434synthesiser for complexity assertions (the blue lines in \autoref{itest1}),
435which can be overridden by the user to increase or decrease accuracy. From the
436assertions, a general purpose deductive platform produces proof obligations
437which in turn can be closed by automatic or interactive provers, ending in a
438proof certificate. Nine proof obligations are generated
439from~\autoref{itest1} (to prove that the loop invariant holds after
440one execution if it holds before, to prove that the whole program execution
441takes at most 1228 cycles, etc.). The CVC3 prover closes all obligations
442in a few seconds on routine commodity hardware.
443%
444\fvset{commandchars=\\\{\}}
445\lstset{morecomment=[s][\color{blue}]{/*@}{*/},
446        moredelim=[is][\color{blue}]{$}{$},
447        moredelim=[is][\color{red}]{|}{|}}
448\begin{figure}[!p]
449\begin{lstlisting}
450|int __cost = 33; int __stack = 5, __stack_max = 5;|
451|void __cost_incr(int incr) { __cost = __cost + incr; }|
452|void __stack_incr(int incr) {
453  __stack = __stack + incr;
454  __stack_max = __stack_max < __stack ? __stack : __stack_max;
455}|
456
457char a[4] = { 3, 2, 7, 252, };
458char treshold = 4;
459
460/*@ behaviour stack_cost:
461      ensures __stack_max <=
462              __max(\old(__stack_max), \old(__stack));
463      ensures __stack == \old(__stack);
464     
465    behaviour time_cost:
466      ensures __cost <= \old(__cost)+1228; */
467int main(void)
468{
469  char j;
470  char *p;
471  int found;
472  |__stack_incr(0); __cost_incr(91);|
473  p = a;
474  found = 0;
475  $__l: /* internal */$
476  /*@ for time_cost: loop invariant
477        __cost <=
478        \at(__cost,__l)+220*(__max(\at(5-j,__l), 0)
479                       -__max (5-j, 0));
480      for stack_cost: loop invariant
481        __stack_max == \at(__stack_max,__l);
482      for stack_cost: loop invariant
483        __stack == \at(__stack,__l);
484      loop variant 4-j; */
485  for (j = 0; j < 4; j++) {
486    |__cost_incr(76);|
487    if (*p <= treshold) {
488      |__cost_incr(144);|
489      found++;
490    } else {
491      |__cost_incr(122);|
492    }
493    p++;
494  }
495  |__cost_incr(37); __stack_incr(-0);|
496  return found;
497}
498\end{lstlisting}
499\caption{The instrumented version of the program in \autoref{test1},
500 with instrumentation added by the CerCo compiler in red and cost invariants
501 added by the CerCo Frama-C plugin in blue. The \texttt{\_\_cost},
502 \texttt{\_\_stack} and \texttt{\_\_stack\_max} variables hold the elapsed time
503in clock cycles and the current and maximum stack usage. Their initial values
504hold the clock cycles spent in initialising the global data before calling
505\texttt{main} and the space required by global data (and thus unavailable for
506the stack).
507}
508\label{itest1}
509\end{figure}
510\section{Main scientific and technical results}
511We will now review the main results that the CerCo project has achieved.
512
513% \emph{Dependent labeling~\cite{?}.} The basic labeling approach assumes a
514% bijective mapping between object code and source code O(1) blocks (called basic
515% blocks). This assumption is violated by many program optimizations (e.g. loop
516% peeling and loop unrolling). It also assumes the cost model computed on the
517% object code to be non parametric: every block must be assigned a cost that does
518% not depend on the state. This assumption is violated by stateful hardware like
519% pipelines, caches, branch prediction units. The dependent labeling approach is
520% an extension of the basic labeling approach that allows to handle parametric
521% cost models. We showed how the method allows to deal with loop optimizations and
522% pipelines, and we speculated about its applications to caches.
523%
524% \emph{Techniques to exploit the induced cost model.} Every technique used for
525% the analysis of functional properties of programs can be adapted to analyse the
526% non-functional properties of the source code instrumented by compilers that
527% implement the labeling approach. In order to gain confidence in this claim, we
528% showed how to implement a cost invariant generator combining abstract
529% interpretation with separation logic ideas \cite{separation}. We integrated
530% everything in the Frama-C modular architecture, in order to compute proof
531% obligations from functional and cost invariants and to use automatic theorem
532% provers on them. This is an example of a new technique that is not currently
533% exploited in WCET analysis. It also shows how precise functional invariants
534% benefits the non-functional analysis too. Finally, we show how to fully
535% automatically analyse the reaction time of Lustre nodes that are first compiled
536% to C using a standard Lustre compiler and then processed by a C compiler that
537% implements the labeling approach.
538
539% \emph{The CerCo compiler.} This is a compiler from a large subset of the C
540% program to 8051/8052 object code,
541% integrating the labeling approach and a static analyser for 8051 executables.
542% The latter can be implemented easily and does not require dependent costs
543% because the 8051 microprocessor is a very simple one, with constant-cost
544% instruction. It was chosen to separate the issue of exact propagation of the
545% cost model from the orthogonal problem of the static analysis of object code
546% that may require approximations or dependent costs. The compiler comes in
547% several versions: some are prototypes implemented directly in OCaml, and they
548% implement both the basic and dependent labeling approaches; the final version
549% is extracted from a Matita certification and at the moment implements only the
550% basic approach.
551
552\subsection{The (basic) labeling approach}
553The labeling approach is the foundational insight that underlies all the developments in CerCo.
554It allows the tracking of the evolution of
555basic blocks during the compilation process in order to propagate the cost model from the
556object code to the source code without losing precision in the process.
557\paragraph{Problem statement.} Given a source program $P$, we want to obtain an
558instrumented source program $P'$,  written in the same programming language, and
559the object code $O$ such that: 1) $P'$ is obtained by inserting into $P$ some
560additional instructions to update global cost information like the amount of
561time spent during execution or the maximal stack space required; 2) $P$ and $P'$ 
562must have the same functional behaviour, i.e.\ they must produce that same output
563and intermediate observables; 3) $P$ and $O$ must have the same functional
564behaviour; 4) after execution and in interesting points during execution, the
565cost information computed by $P'$ must be an upper bound of the one spent by $O$ 
566to perform the corresponding operations (soundness property); 5) the difference
567between the costs computed by $P'$ and the execution costs of $O$ must be
568bounded by a program-dependent constant (precision property).
569
570\paragraph{The labeling software components.} We solve the problem in four
571stages \cite{labeling}, implemented by four software components that are used
572in sequence.
573
574The first component labels the source program $P$ by injecting label emission
575statements in appropriate positions to mark the beginning of basic blocks.
576% The
577% set of labels with their positions is called labeling.
578The syntax and semantics
579of the source programming language is augmented with label emission statements.
580The statement ``EMIT $\ell$'' behaves like a NOP instruction that does not affect the
581program state or control flow, but its execution is observable.
582% Therefore the observables of a run of a program becomes a stream
583% of label emissions: $\ell_1,\ldots,\ell_n$, called the program trace. We clarify the conditions
584% that the labeling must respect later.
585
586The second component is a labeling preserving compiler. It can be obtained from
587an existing compiler by adding label emission statements to every intermediate
588language and by propagating label emission statements during compilation. The
589compiler is correct if it preserves both the functional behaviour of the program
590and the traces of observables.
591% We may also ask that the function that erases the cost
592% emission statements commute with compilation. This optional property grants that
593% the labeling does not interfere with the original compiler behaviour. A further
594% set of requirements will be added later.
595
596The third component is a static object code analyser. It takes as input a labeled
597object code and it computes the scope of each of its label emission statements,
598i.e.\ the tree of instructions that may be executed after the statement and before
599a new label emission is encountered.
600It is a tree and not a sequence because the scope
601may contain a branching statement. In order to grant that such a finite tree
602exists, the object code must not contain any loop that is not broken by a label
603emission statement. This is the first requirement of a sound labeling. The
604analyser fails if the labeling is unsound. For each scope, the analyser
605computes an upper bound of the execution time required by the scope, using the
606maximum of the costs of the two branches in case of a conditional statement.
607Finally, the analyser computes the cost of a label by taking the maximum of the
608costs of the scopes of every statement that emits that label.
609
610The fourth and last component takes in input the cost model computed at step 3
611and the labelled code computed at step 1. It outputs a source program obtained
612by replacing each label emission statement with a statement that increments the
613global cost variable with the cost associated to the label by the cost model. 
614The obtained source code is the instrumented source code.
615
616\paragraph{Correctness.} Requirements 1, 2 and 3 of the problem statement
617obviously hold, with 2 and 3 being a consequence of the definition of a correct
618labeling preserving compiler. It is also obvious that the value of the global
619cost variable of an instrumented source code is at any time equal to the sum of
620the costs of the labels emitted by the corresponding labelled code. Moreover,
621because the compiler preserves all traces, the sum of the costs of the labels
622emitted in the source and target labelled code are the same. Therefore, to
623satisfy the fourth requirement, we need to grant that the time taken to execute
624the object code is equal to the sum of the costs of the labels emitted by the
625object code. We collect all the necessary conditions for this to happen in the
626definition of a sound labeling: a) all loops must be broken by a cost emission
627statement;  b) all program instructions must be in the scope of some cost
628emission statement. To satisfy also the fifth requirement, additional
629requirements must be imposed on the object code labeling to avoid all uses of
630the maximum in the cost computation: the labeling is precise if every label is
631emitted at most once and both branches of a conditional jump start with a label
632emission statement.
633
634The correctness and precision of the labeling approach only rely on the
635correctness and precision of the object code labeling. The simplest
636way to achieve them is to impose correctness and precision
637requirements also on the initial labeling produced by the first software
638component, and to ask the compiler to preserve these
639properties too. The latter requirement imposes serious limitations on the
640compilation strategy and optimizations: the compiler may not duplicate any code
641that contains label emission statements, like loop bodies. Therefore several
642loop optimisations like peeling or unrolling are prevented. Moreover, precision
643of the object code labeling is not sufficient \emph{per se} to obtain global
644precision: we implicitly assumed that a precise constant cost can be assigned
645to every instruction. This is not possible in
646the presence of stateful hardware whose state influences the cost of operations,
647like pipelines and caches. In the next subsection we will see an extension of the
648basic labeling approach to cover this situation.
649
650The labeling approach described in this section can be applied equally well and
651with minor modifications to imperative and functional languages
652\cite{labeling2}. We have tested it on a simple imperative language without
653functions (a `while' language), on a subset of C and on two compilation chains for
654a purely functional higher-order language. The two main changes required to deal
655with functional languages are: 1) because global variables and updates are not
656available, the instrumentation phase produces monadic code to `update' the
657global costs; 2) the requirements for a sound and precise labeling of the
658source code must be changed when the compilation is based on CPS translations.
659In particular, we need to introduce labels emitted before a statement is
660executed and also labels emitted after a statement is executed. The latter capture
661code that is inserted by the CPS translation and that would escape all label
662scopes.
663
664% Brian: one of the reviewers pointed out that standard Prolog implementations
665% do have some global state that apparently survives backtracking and might be
666% used.  As we haven't experimented with this, I think it's best to elide it
667% entirely.
668
669% Phases 1, 2 and 3 can be applied as well to logic languages (e.g. Prolog).
670% However, the instrumentation phase cannot: in standard Prolog there is no notion
671% of (global) variable whose state is not retracted during backtracking.
672% Therefore, the cost of executing computations that are later backtracked would
673% not be correctly counted in. Any extension of logic languages with
674% non-backtrackable state could support our labeling approach.
675
676\subsection{Dependent labeling}
677The core idea of the basic labeling approach is to establish a tight connection
678between basic blocks executed in the source and target languages. Once the
679connection is established, any cost model computed on the object code can be
680transferred to the source code, without affecting the code of the compiler or
681its proof. In particular, it is immediate that we can also transport cost models
682that associate to each label functions from hardware state to natural numbers.
683However, a problem arises during the instrumentation phase that replaces cost
684emission statements with increments of global cost variables. The global cost
685variable must be incremented with the result of applying the function associated
686to the label to the hardware state at the time of execution of the block.
687The hardware state comprises both the functional state that affects the
688computation (the value of the registers and memory) and the non-functional
689state that does not (the pipeline and cache contents for example). The former is
690in correspondence with the source code state, but reconstructing the
691correspondence may be hard and lifting the cost model to work on the source code
692state is likely to produce cost expressions that are too complex to understand and reason about.
693Luckily enough, in all modern architectures the cost of executing single
694instructions is either independent of the functional state or the jitter---the
695difference between the worst and best case execution times---is small enough
696to be bounded without losing too much precision. Therefore we can concentrate
697on dependencies over the `non-functional' parts of the state only.
698
699The non-functional state has no correspondence in the high level state and does
700not influence the functional properties. What can be done is to expose the
701non-functional state in the source code. We present here the basic
702intuition in a simplified form: the technical details that allow us to handle the
703general case are more complex and can be found in~\cite{paolo}. We add
704to the source code an additional global variable that represents the
705non-functional state and another one that remembers the last labels emitted. The
706state variable must be updated at every label emission statement, using an
707update function which is computed during static analysis too. The update
708function associates to each label a function from the recently emitted labels
709and old state to the new state. It is computed composing the semantics of every
710instruction in a basic block and restricting it to the non-functional part of
711the state.
712
713Not all the details of the non-functional state needs to be exposed, and the
714technique works better when the part of state that is required can be summarized
715in a simple data structure. For example, to handle simple but realistic
716pipelines it is sufficient to remember a short integer that encodes the position
717of bubbles (stuck instructions) in the pipeline. In any case, it is not necessary
718for the user to understand the meaning of the state to reason over the properties
719of the
720program. Moreover, at any moment the user, or the invariant generator tools that
721analyse the instrumented source code produced by the compiler, can decide to
722trade precision of the analysis for simplicity by approximating the parametric
723cost with safe non parametric bounds. Interestingly, the functional analysis of
724the code can determine which blocks are executed more frequently in order to
725approximate more aggressively the ones that are executed less.
726
727Dependent labeling can also be applied to allow the compiler to duplicate
728blocks that contain labels (e.g. in loop optimisations)~\cite{paolo}. The effect is to assign
729a different cost to the different occurrences of a duplicated label. For
730example, loop peeling turns a loop into the concatenation of a copy of the loop
731body (that executes the first iteration) with the conditional execution of the
732loop (for the successive iterations). Because of further optimisations, the two
733copies of the loop will be compiled differently, with the first body usually
734taking more time.
735
736By introducing a variable that keeps track of the iteration number, we can
737associate to the label a cost that is a function of the iteration number. The
738same technique works for loop unrolling without modifications: the function will
739assign a cost to the even iterations and another cost to the odd ones. The
740actual work to be done consists of introducing within the source code, and for each
741loop, a variable that counts the number of iterations. The loop optimisation code
742that duplicate the loop bodies must also modify the code to propagate correctly
743the update of the iteration numbers. The technical details are more complex and
744can be found in the CerCo reports and publications. The implementation, however,
745is quite simple and the changes to a loop optimising compiler are minimal.
746
747\subsection{Techniques to exploit the induced cost model}
748We review the cost synthesis techniques developed in the project.
749The starting hypothesis is that we have a certified methodology to annotate
750blocks in the source code with constants which provide a sound and sufficiently
751precise upper bound on the cost of executing the blocks after compilation to
752object code.
753
754The principle that we have followed in designing the cost synthesis tools is
755that the synthetic bounds should be expressed and proved within a general
756purpose tool built to reason on the source code. In particular, we rely on the
757Frama-C tool to reason on C code and on the Coq proof-assistant to reason on
758higher-order functional programs.
759
760This principle entails that: 1)
761The inferred synthetic bounds are indeed correct as long as the general purpose
762tool is; 2) there is no limitation on the class of programs that can be handled
763as long as the user is willing to carry on an interactive proof.
764
765Of course, automation is desirable whenever possible. Within this framework,
766automation means writing programs that give hints to the general purpose tool.
767These hints may take the form, say, of loop invariants/variants, of predicates
768describing the structure of the heap, or of types in a light logic. If these
769hints are correct and sufficiently precise the general purpose tool will produce
770a proof automatically, otherwise, user interaction is required.
771
772\paragraph{The Cost plugin and its application to the Lustre compiler.}
773Frama-C \cite{framac} is a set of analysers for C programs with a
774specification language called ACSL. New analyses can be dynamically added
775via a plugin system. For instance, the Jessie plugin allows deductive
776verification of C programs with respect to their specification in ACSL, with
777various provers as back-end tools.
778We developed the CerCo Cost plugin for the Frama-C platform as a proof of
779concept of an automatic environment exploiting the cost annotations produced by
780the CerCo compiler. It consists of an OCaml program which essentially
781takes the following actions: 1) it receives as input a C program, 2) it
782applies the CerCo compiler to produce a related C program with cost annotations,
7833) it applies some heuristics to produce a tentative bound on the cost of
784executing the C functions of the program as a function of the value of their
785parameters, 4) the user can then call the Jessie plugin to discharge the
786related proof obligations.
787In the following we elaborate on the soundness of the framework and the
788experiments we performed with the Cost tool on the C programs produced by a
789Lustre compiler.
790
791\paragraph{Soundness.} The soundness of the whole framework depends on the cost
792annotations added by the CerCo compiler, the synthetic costs produced by the
793cost plugin, the verification conditions (VCs) generated by Jessie, and the
794external provers discharging the VCs. The synthetic costs being in ACSL format,
795Jessie can be used to verify them. Thus, even if the added synthetic costs are
796incorrect (relatively to the cost annotations), the process as a whole is still
797correct: indeed, Jessie will not validate incorrect costs and no conclusion can
798be made about the WCET of the program in this case. In other terms, the
799soundness does not really depend on the action of the cost plugin, which can in
800principle produce any synthetic cost. However, in order to be able to actually
801prove a WCET of a C function, we need to add correct annotations in a way that
802Jessie and subsequent automatic provers have enough information to deduce their
803validity. In practice this is not straightforward even for very simple programs
804composed of branching and assignments (no loops and no recursion) because a fine
805analysis of the VCs associated with branching may lead to a complexity blow up.
806\paragraph{Experience with Lustre.} Lustre is a data-flow language for programming
807synchronous systems, with a compiler which targets C. We designed a
808wrapper for supporting Lustre files.
809The C function produced by the compiler implements the step function of the
810synchronous system and computing the WCET of the function amounts to obtain a
811bound on the reaction time of the system. We tested the Cost plugin and the
812Lustre wrapper on the C programs generated by the Lustre compiler. For programs
813consisting of a few hundred lines of code, the cost plugin computes a WCET and Alt-
814Ergo is able to discharge all VCs automatically.
815
816\paragraph{Handling C programs with simple loops.}
817The cost annotations added by the CerCo compiler take the form of C instructions
818that update by a constant a fresh global variable called the cost variable.
819Synthesizing a WCET of a C function thus consists in statically resolving an
820upper bound of the difference between the value of the cost variable before and
821after the execution of the function, i.e. find in the function the instructions
822that update the cost variable and establish the number of times they are passed
823through during the flow of execution. In order to do the analysis the plugin
824makes the following assumptions on the programs:
8251) there are no recursive functions;
8262) every loop must be annotated with a variant. The variants of `for' loops are
827automatically inferred.
828
829The plugin proceeds as follows.
830First the call graph of the program is computed.
831Then the computation of the cost of the function is performed by traversing its
832control flow graph. If the function $f$ calls the function $g$ 
833then the function $g$ 
834is processed before the function $f$. The cost at a node is the maximum of the
835costs of the successors.
836In the case of a loop with a body that has a constant cost for every step of the
837loop, the cost is the product of the cost of the body and of the variant taken
838at the start of the loop.
839In the case of a loop with a body whose cost depends on the values of some free
840variables, a fresh logic function $f$ is introduced to represent the cost of the
841loop in the logic assertions. This logic function takes the variant as a first
842parameter. The other parameters of $f$ are the free variables of the body of the
843loop. An axiom is added to account the fact that the cost is accumulated at each
844step of the loop.
845The cost of the function is directly added as post-condition of the function.
846
847The user can influence the annotation by two different means:
8481) by using more precise variants;
8492) by annotating functions with cost specifications. The plugin will use this cost
850for the function instead of computing it.
851\paragraph{C programs with pointers.}
852When it comes to verifying programs involving pointer-based data structures,
853such as linked lists, trees, or graphs, the use of traditional first-order logic
854to specify, and of SMT solvers to verify, shows some limitations. Separation
855logic~\cite{separation} is an elegant alternative. It is a program logic
856with a new notion of conjunction to express spatial heap separation. Bobot has
857recently introduced automatically generated separation
858predicates to simulate separation logic reasoning in the Jessie plugin where the specification language, the verification condition
859generator, and the theorem provers were not designed with separation logic in
860mind. CerCo's plugin can exploit the separation predicates to automatically
861reason on the cost of execution of simple heap manipulation programs such as an
862in-place list reversal.
863
864\subsection{The CerCo compiler}
865In CerCo we have developed a certain number of cost preserving compilers based
866on the labeling approach. Excluding an initial certified compiler for a `while'
867language, all remaining compilers target realistic source languages---a pure
868higher order functional language and a large subset of C with pointers, gotos
869and all data structures---and real world target processors---MIPS and the
870Intel 8051/8052 processor family. Moreover, they achieve a level of optimisation
871that ranges from moderate (comparable to GCC level 1) to intermediate (including
872loop peeling and unrolling, hoisting and late constant propagation). The so
873called \emph{Trusted CerCo Compiler} is the only one that was implemented in the
874interactive theorem prover Matita~\cite{matita} and its costs certified. The code distributed
875is extracted OCaml code from the Matita implementation. In the rest of
876this section we will only focus on the Trusted CerCo Compiler, that only targets
877the C language and the 8051/8052 family, and that does not implement any
878advanced optimisations yet. Its user interface, however, is the same as the one
879of the other versions, in order to trade safety with additional performances. In
880particular, the Frama-C CerCo plugin can work without recompilation with all
881compilers.
882
883The 8051/8052 microprocessor is a very simple one, with constant-cost
884instructions. It was chosen to separate the issue of exact propagation of the
885cost model from the orthogonal problem of the static analysis of object code
886that may require approximations or dependent costs.
887
888The (trusted) CerCo compiler implements the following optimisations: cast
889simplification, constant propagation in expressions, liveness analysis driven
890spilling of registers, dead code elimination, branch displacement, and tunneling.
891The two latter optimisations are performed by our optimising assembler
892\cite{correctness}. The back-end of the compiler works on three address
893instructions, preferred to static single assignment code for the simplicity of
894the formal certification.
895
896The CerCo compiler is loosely based on the CompCert compiler \cite{compcert}, a
897recently developed certified compiler from C to the PowerPC, ARM and x86
898microprocessors. Contrary to CompCert, both the CerCo code and its
899certification are open source. Some data structures and language definitions for
900the front-end are directly taken from CompCert, while the back-end is a redesign
901of a compiler from Pascal to MIPS used by Fran\c{c}ois Pottier for a course at the
902Ecole Polytechnique.
903
904The main peculiarities of the CerCo compiler are the following.
905\begin{enumerate}
906\item All the intermediate languages include label emitting instructions to
907implement the labeling approach, and the compiler preserves execution traces.
908\item Traditionally compilers target an assembly language with additional
909macro-instructions to be expanded before assembly; for CerCo we need to go all
910the way down to object code in order to perform the static analysis. Therefore
911we integrated also an optimising assembler and a static analyser.
912\item In order to avoid implementing a linker and a loader, we do not support separate
913compilation and external calls. Adding them is a transparent
914process to the labeling approach and should create no unknown problem.
915\item We target an 8-bit processor, in contrast to CompCert's 32-bit targets. Targeting an 8-bit processor requires
916several changes and increased code size, but it is not fundamentally more
917complex. The proof of correctness, however, becomes much harder.
918\item We target a microprocessor that has a non uniform memory model, which is
919still often the case for microprocessors used in embedded systems and that is
920becoming common again in multi-core processors. Therefore the compiler has to
921keep track of data and it must move data between memory regions in the proper
922way. Moreover the size of pointers to different regions is not uniform. An
923additional difficulty was that the space available for the stack in internal
924memory in the 8051 is tiny, allowing only a minor number of nested calls. To
925support full recursion in order to test the CerCo tools also on recursive
926programs, the compiler implements a stack in external memory.
927\end{enumerate}
928
929\subsection{Formal certification of the CerCo compiler}
930We implemented the
931CerCo compiler in the interactive theorem prover Matita and have formally
932certified that the cost model induced on the source code correctly and precisely
933predicts the object code behaviour. We actually induce two cost models, one for
934time and one for stack space consumption. We show the correctness of the prediction
935only for those programs that do not exhaust the available machine stack space, a
936property that---thanks to the stack cost model---we can statically analyse on the
937source code using our Frama-C tool. The preservation of functional properties we
938take as an assumption, not itself formally proved in CerCo.  Other projects have
939already certified the preservation of functional semantics in similar compilers,
940and we have not attempted to directly repeat that work. In order to complete the
941proof for non-functional properties, we have introduced a new semantics for
942programming languages based on a new kind of structured observables with the
943relative notions of forward similarity and the study of the intentional
944consequences of forward similarity. We have also introduced a unified
945representation for back-end intermediate languages that was exploited to provide
946a uniform proof of forward similarity.
947
948The details on the proof techniques employed
949and
950the proof sketch can be collected from the CerCo deliverables and papers.
951In this section we will only hint at the correctness statement, which turned
952out to be more complex than what we expected at the beginning.
953
954\paragraph{The statement.}
955Real time programs are often reactive programs that loop forever responding to
956events (inputs) by performing some computation followed by some action (output)
957and the return to the initial state. For looping programs the overall execution
958time does not make sense. The same happens for reactive programs that spend an
959unpredictable amount of time in I/O. What is interesting is the reaction time
960that measure the time spent between I/O events. Moreover, we are interested in
961predicting and ruling out programs that crash running out of space on a certain
962input.
963Therefore we need to look for a statement that talks about sub-runs of a
964program. The most natural statement is that the time predicted on the source
965code and spent on the object code by two corresponding sub-runs are the same.
966The problem to solve to make this statement formal is how to identify the
967corresponding sub-runs and how to single out those that are meaningful.
968The solution we found is based on the notion of measurability. We say that a run
969has a \emph{measurable sub-run} when both the prefix of the sub-run and the
970sub-run do not exhaust the stack space, the number of function calls and returns
971in the sub-run is the same, the sub-run does not perform any I/O and the sub-run
972starts with a label emission statement and ends with a return or another label
973emission statements. The stack usage is estimated using the stack usage model
974that is computed by the compiler.
975
976The statement that we formally proved is: for each C run with a measurable
977sub-run, there exists an object code run with a sub-run, such that the observables
978of the pairs of prefixes and sub-runs are the same and the time spent by the
979object code in the sub-run is the same as the one predicted on the source code
980using the time cost model generated by the compiler.
981We briefly discuss the constraints for measurability. Not exhausting the stack
982space is a clear requirement of meaningfulness of a run, because source programs
983do not crash for lack of space while object code ones do. The balancing of
984function calls and returns is a requirement for precision: the labeling approach
985allows the scope of label emission statements to extend after function calls to
986minimize the number of labels. Therefore a label pays for all the instructions
987in a block, excluding those executed in nested function calls. If the number of
988calls/returns is unbalanced, it means that there is a call we have not returned
989to that could be followed by additional instructions whose cost has already been
990taken in account. To make the statement true (but less precise) in this
991situation, we could only say that the cost predicted on the source code is a
992safe bound, not that it is exact. The last condition on the entry/exit points of
993a run is used to identify sub-runs whose code correspond to a sequence of blocks
994that we can measure precisely. Any other choice would start or end the run in the
995middle of a block and we would be forced again to weaken the statement taking as
996a bound the cost obtained counting in all the instructions that precede the
997starting one in the block, or follow the final one in the block.
998I/O operations can be performed in the prefix of the run, but not in the
999measurable sub-run. Therefore we prove that we can predict reaction times, but
1000not I/O times, as it should be.
1001
1002\section{Conclusions and future work}
1003
1004All the CerCo software and deliverables can be found on the CerCo homepage at~\url{http://cerco.cs.unibo.it}.
1005
1006The results obtained so far are encouraging and provide evidence that
1007it is possible to perform static time and space analysis at the source level
1008without losing accuracy, reducing the trusted code base and reconciling the
1009study of functional and non-functional properties of programs. The
1010techniques introduced seem to be scalable, cover both imperative and
1011functional languages and are compatible with every compiler optimisation
1012considered by us so far.
1013
1014To prove that compilers can keep track of optimisations
1015and induce a precise cost model on the source code, we targeted a simple
1016architecture that admits a cost model that is execution history independent.
1017The most important future work is dealing with hardware architectures
1018characterized by history dependent stateful components, like caches and
1019pipelines. The main issue consists in assigning a parametric, dependent cost
1020to basic blocks that can be later transferred by the labeling approach to
1021the source code and represented in a meaningful way to the user. The dependent
1022labeling approach that we have studied seems a promising tool to achieve
1023this goal, but the cost model generated for a realistic processor could be too
1024large and complex to be exposed in the source code. Further study is required
1025to evaluate the technique on a realistic processor and to introduce early
1026approximations of the cost model to make the technique feasible.
1027
1028Examples of further future work consist in improving the cost invariant
1029generator algorithms and the coverage of compiler optimizations, in combining
1030the labeling approach with the type and effect discipline of~\cite{typeffects}
1031to handle languages with implicit memory management, and in experimenting with
1032our tools in early development phases. Some larger case study is also necessary
1033to evaluate the CerCo's prototype on realistic, industrial-scale programs.
1034
1035% \bibliographystyle{splncs}
1036\bibliography{fopara13}
1037% \begin{thebibliography}{19}
1038%
1039% \bibitem{survey} \textbf{A Survey of Static Program Analysis Techniques}
1040% W.~W\"ogerer. Technical report. Technische Universit\"at Wien 2005
1041%
1042% \bibitem{cerco} \textbf{Certified Complexity}. R.M. Amadio, A. Asperti, N. Ayache,
1043% B. Campbell, D. P. Mulligan, R. Pollack, Y. Regis-Gianas, C. Sacerdoti Coen, I.
1044% Stark, in Procedia Computer Science, Volume 7, 2011, Proceedings of the 2 nd
1045% European Future Technologies Conference and Exhibition 2011 (FET 11), 175-177.
1046%
1047% \bibitem{labeling} \textbf{Certifying and Reasoning on Cost Annotations in C
1048% Programs}, N.  Ayache, R.M. Amadio, Y.R\'{e}gis-Gianas, in Proc. FMICS, Springer
1049% LNCS
1050% 7437: 32-46, 2012.
1051% %, DOI:10.1007/978-3-642-32469-7\_3.
1052%
1053% \bibitem{labeling2} \textbf{Certifying and reasoning on cost annotations of
1054% functional programs}.
1055% R.M. Amadio, Y. R\'{e}gis-Gianas. Proceedings of the Second international conference
1056% on Foundational and Practical Aspects of Resource Analysis FOPARA 2011 Springer
1057% LNCS 7177:72-89, 2012.
1058%
1059% \bibitem{compcert} \textbf{Formal verification of a realistic compiler}. X. Leroy,  In Commun. ACM 52(7), 107–115, 2009.
1060%
1061% \bibitem{framac} \textbf{Frama-C user manual}. L. Correnson, P. Cuoq, F. Kirchner, V. Prevosto, A. Puccetti, J. Signoles,
1062% B. Yakobowski. in CEA-LIST, Software Safety Laboratory, Saclay, F-91191,
1063% \url{http://frama-c.com/}.
1064%
1065% \bibitem{paolo} \textbf{Indexed Labels for Loop Iteration Dependent Costs}. P.
1066% Tranquilli, in Proceedings of the 11th International Workshop on Quantitative
1067% Aspects of Programming Languages and Systems (QAPL 2013), Rome, 23rd-24th March
1068% 2013, Electronic Proceedings in Theoretical Computer Science, to appear in 2013.
1069%
1070% \bibitem{separation} \textbf{Intuitionistic reasoning about shared mutable data
1071% structure} J.C. Reynolds. In Millennial Perspectives in Computer Science,
1072% pages 303–321, Houndsmill, Hampshire, 2000. Palgrave.
1073%
1074% \bibitem{lustre} \textbf{LUSTRE: a declarative language for real-time
1075% programming}
1076% P. Caspi, D. Pilaud, N. Halbwachs, J.A. Plaice. In Proceedings of
1077% the 14th ACM SIGACT-SIGPLAN symposium on Principles of programming languages ACM
1078% 1987.
1079%
1080% \bibitem{correctness} \textbf{On the correctness of an optimising assembler for
1081% the intel MCS-51 microprocessor}.
1082%   D. P. Mulligan, C. Sacerdoti Coen. In Proceedings of the Second
1083% international conference on Certified Programs and Proofs, Springer-Verlag 2012.
1084%
1085% \bibitem{proartis} \textbf{PROARTIS: Probabilistically Analysable Real-Time
1086% Systems}, F.J. Cazorla, E. Qui\~{n}ones, T. Vardanega, L. Cucu, B. Triquet, G.
1087% Bernat, E. Berger, J. Abella, F. Wartel, M. Houston, et al., in ACM Transactions
1088% on Embedded Computing Systems, 2012.
1089%
1090% \bibitem{embounded} \textbf{The EmBounded project (project paper)}. K. Hammond,
1091% R. Dyckhoff, C. Ferdinand, R. Heckmann, M. Hofmann, H. Loidl, G. Michaelson, J.
1092% Serot, A. Wallace, in Trends in Functional Programming, Volume 6, Intellect
1093% Press, 2006.
1094%
1095% \bibitem{matita}
1096% \textbf{The Matita Interactive Theorem Prover}.
1097% A. Asperti, C. Sacerdoti Coen, W. Ricciotti, E. Tassi.
1098% 23rd International Conference on Automated Deduction, CADE 2011.
1099%
1100% \bibitem{typeffects} \textbf{The Type and Effect Discipline}. J.-P. Talpin,
1101%  P. Jouvelot.
1102%   In Proceedings of the Seventh Annual Symposium on Logic in Computer Science
1103% (LICS '92), Santa Cruz, California, USA, June 22-25, 1992.
1104% IEEE Computer Society 1992.
1105%
1106% \bibitem{stateart} \textbf{The worst-case execution-time problem overview of
1107% methods
1108% and survey of tools.} R. Wilhelm et al., in  ACM Transactions on Embedded
1109% Computing Systems, 7:1–53, May 2008.
1110%
1111% %\bibitem{proartis2} \textbf{A Cache Design for Probabilistic Real-Time
1112% % Systems}, L. Kosmidis, J. Abella, E. Quinones, and F. Cazorla, in Design,
1113% % Automation, and Test in Europe (DATE), Grenoble, France, 03/2013.
1114%
1115% \end{thebibliography}
1116
1117
1118%\bibliography{fopara13.bib}
1119
1120\end{document}
Note: See TracBrowser for help on using the repository browser.